Hybrid-COVID: a novel hybrid 2D/3D CNN based on cross-domain adaptation approach for COVID-19 screening from chest X-ray images

Physical and Engineering Sciences in Medicine (2020) 43:1415–1431 https://doi.org/10.1007/s13246-020-00957-1 SCIENTIFIC PAPER Hybrid‑COVID: a novel hybrid 2D/3D CNN based on cross‑domain adaptation approach for COVID‑19 screening from chest X‑ray images Khaled Bayoudh1 · Fayçal Hamdaoui2 · Abdellatif Mtibaa1 Received: 19 August 2020 / Accepted: 2 December 2020 / Published online: 10 December 2020 © Australasian College of Physical Scientists and Engineers in Medicine 2020 Abstract The novel Coronavirus disease (COVID-19), which first appeared at the end of December 2019, continues to spread rapidly in most countries of the world. Respiratory infections occur primarily in the majority of patients treated with COVID-19. In light of the growing number of COVID-19 cases, the need for diagnostic tools to identify COVID-19 infection at early stages is of vital importance. For decades, chest X-ray (CXR) technologies have proven their ability to accurately detect respiratory diseases. More recently, with the availability of COVID-19 CXR scans, deep learning algorithms have played a critical role in the healthcare arena by allowing radiologists to recognize COVID-19 patients from their CXR images. However, the majority of screening methods for COVID-19 reported in recent studies are based on 2D convolutional neural networks (CNNs). Although 3D CNNs are capable of capturing contextual information compared to their 2D counterparts, their use is limited due to their increased computational cost (i.e. requires much extra memory and much more computing power). In this study, a transfer learning-based hybrid 2D/3D CNN architecture for COVID-19 screening using CXRs has been developed. The proposed architecture consists of the incorporation of a pre-trained deep model (VGG16) and a shallow 3D CNN, combined with a depth-wise separable convolution layer and a spatial pyramid pooling module (SPP). Specifically, the depth-wise separable convolution helps to preserve the useful features while reducing the computational burden of the model. The SPP module is designed to extract multi-level representations from intermediate ones. Experimental results show that the proposed framework can achieve reasonable performances when evaluated on a collected dataset (3 classes to be predicted: COVID-19, Pneumonia, and Normal). Notably, it achieved a sensitivity of 98.33%, a specificity of 98.68% and an overall accuracy of 96.91% Keywords COVID-19 · Chest X-ray · Hybrid 2D/3D CNN · Deep learning · Pneumonia Introduction The novel Coronavirus (COVID-19), which originated in Wuhan, China at the end of 2019, has become a serious threat worldwide to public health [1]. Accordingly, this * Khaled Bayoudh 1 Electrical Department, National Engineering School of Monastir (ENIM), Laboratory of Electronics and Micro‑electronics (LR99ES30), Faculty of Sciences of Monastir (FSM), University of Monastir, Monastir, Tunisia 2 Electrical Department, National Engineering School of Monastir (ENIM), Laboratory of Control, Electrical Systems and Environment (LASEE), National Engineering School of Monastir (ENIM), University of Monastir, Monastir, Tunisia pandemic stands as a global health emergency. The continuing spread of the COVID-19 pandemic, which claims a large number of victims and serious infections every day, affecting several territories, such as the United States, Italy, Spain, etc., makes its treatment increasingly challenging. According to the World Health Organization (WHO), COVID-19 has a considerable impact on human beings. Till November 2020, the number of confirmed COVID-19 cases in the affected countries reached 58,900,547, of which 1,393,305 have died, reported to WHO. The majority of infected patients suffer from mild to moderate respiratory issues, the severity of which changes over time [2]. The disease is highly contagious because it can spread rapidly from infected people to healthy people through micron-size droplets from the nose and oral cavity or close contact between infected and uninfected people; COVID-19 has a reproduction rate of 3 or 13 Vol.:(0123456789) 1416 more, meaning that on average 3 or more people can be infected per COVID-19 case [3]. Currently, the most recognized strategy to fight the pandemic involves intensified screening for the infection [4]. To slow down its spread, millions of people still need to be screened over time. The real-time reverse transcriptase-polymerase chain reaction (RT-PCR) test can detect and screen the presence of the virus with a high-level of sensitivity. This kind of test has often been used as the main screening process for COVID-19 by directly identifying the existence of the virus. Despite this advantage, RT-PCR suffers from some drawbacks. For example, some centers have a limited number of RT-PCR kits available and there is a high risk of receiving false-negative RT-PCR results over time [5]. Moreover, it is very time-consuming and expensive especially when the collected specimens need to be processed by external specialized laboratories. Hence, this makes it very difficult to perform RT-PCR testing for a large number of suspected patients in as short a time as possible. To address these shortcomings, researchers have developed an important complement to RT-PCR tests by detecting COVID-19 from chest computed tomography (CT) scans [6–9]. This will help reduce treatment delays and patient isolation. In [6], Ai et al. concluded that chest CT scans were more sensitive for the diagnosis of COVID-19 than traditional RT-PCR, although it should be noted this work was published very early in the pandemic. In [7], Dangis et al. showed that the chest CT scanner has considerable performance in diagnosing COVID-19 in terms of speed, specificity, and sensitivity. A comparative study of 51 patients showed that chest CT scans were very sensitive (50/51 patients) for screening for COVID-19 disease [8]. In a study by Bai et al. [9], a total of 219 COVID-19 positive patients detected by both RT-PCR and chest CT were conducted. The researchers concluded quantitatively that the distinction Physical and Engineering Sciences in Medicine (2020) 43:1415–1431 of COVID-19 from other types of pneumonia on CT chest scans was more sensitive than RT-PCR tests. Regardless of these strengths, CT scans still have some limitations. Among these is the fact that screening for COVID-19 generally takes much longer to perform. Moreover, high-quality scanners are very expensive and patients have to deal with much higher radiation doses [10]. In contrast, chest X-rays (CXRs) are one of the most common large-scale medical imaging techniques that have proven to be effective in speeding-up and facilitating the screening of the pandemic [11, 12]. These techniques allow the early detection and tracking of the virus. Figure 1 shows CXRs of an older man patient from Wuhan, China [13]. These samples indicate the progression of lung consolidation from days 0, 4, and 7, respectivel (...truncated)